Research On Vehicle Detection Algorithm In Complex Traffic Environment

With the rapid development of artificial intelligence technology,the performance of the core visual perception module as the "eye" of the autonomous car has also been rapidly improved.In real complex traffic scenarios,vehicle detection has become one of the most important issues in the field of computer vision due to the variability of the view angle of the vehicle,the mutual occlusion and the complexity of the apparent form.Current vehicle detection methods are mainly divided into traditional machine learning and deep learning based methods.The traditional machine learning method is mainly divided into two separate parts: feature extraction and classifier decision.It is impossible to collaboratively optimize the information between the two parts,but it has good interpretability.The deep learning method uses the convolutional neural network to complete feature extraction and classification decision simultaneously,and realizes the application-oriented end-to-end learning mode.Although the performance of current vehicle detection algorithms has made great progress,when the road vehicles are in a complex traffic environment where mutual occlusion and dynamic view angle changes frequently occur,the vehicle detection performance will be significantly reduced.Aiming at the above problems,this paper studies the multi-view vehicle detection algorithm based on hybrid difference features and the vehicle detection method based on multi-layer convolutional neural network from the two perspectives of traditional machine learning and deep learning methods,which significantly improves the performance of existing algorithms.The main work is summarized as follows:(1)Conventional multi-view vehicle detectors rely on training multiple singleview object detectors and the discriminating ability to extract the vehicle characterization features directly related to detection performance.This paper proposes a multi-view vehicle detection system based on hybrid pixel differential features,which has high detection performance and real-time running speed.Firstly,based on the three-dimensional information,aspect ratio and occlusion level of the vehicle,the spectral clustering algorithm is used to generate different subcategories of the vehicle,which reduces the model complexity of the single-view object detector.Secondly,a feature extraction based on hybrid differential feature is proposed for each vehicle subcategory.The characteristics of the vehicle subcategories indicate that the discriminative ability of the traditional integral channel features is improved.Finally,a cost-sensitive multiclass classifier is used for feature selection to achieve real-time running speed.The method in this paper is validated on the KITTI dataset.The experimental results show that the proposed method can obtain the best performance in the traditional handcrafted feature-based method.(2)For the shortcoming that the region proposal network(RPN)in the classical Faster R-CNN object detection framework only relies on single-scale feature layer,this paper respectively improves the multi-scale feature fusion and the structure of RPN network.For the scale problem,the feature pyramid network is added in the convolution process,and the multi-scale feature is used for classification and bounding box regression.However,the experimental results show that the performance improvement of the vehicle detection is not obvious.According to the anchor location problem,the structural parameters of the RPN network are modified,which makes the RPN network more precise for the location of the vehicle.The results of the method on the KITTI dataset show that the average accuracy of the improved method is higher than the Faster R-CNN algorithm.4.5%.(3)Aiming at the problem of the complementarity of speed and performance of the one-stage and two-stage object detection frameworks,a one-stage and two-stage syncretic object detection algorithm is proposed,which not only has the high detection accuracy of the two-stage method,but also obtains faster detection speed.In this paper,adaptive anchor is used in the anchor refinement module to make the detection framework have better generalization ability.Experimental results on KITTI dataset show that the mAP of the proposed method is 13.5% higher than that of the classical Faster R-CNN method.